My advice: write your own English prose, and try not to let "LLM-speak" leak into your documentation when using them to edit. Ironically, LLMs just plain suck at writing English, like they're incredibly overfit on marketing copy and press releases. I hope someone is working on this, or at least cares about the problem, because that would make this brave new world palatable for reading.
They really don't, if you actually bother prompting them. Give them a voice sample, and tell them to match the tone, and you already get something 10x better. Have them revise with a list of common writing problems - not just common LLM patterns, but guidelines for writing better - and you get rid of more.
Properly prompted, an LLM writes far better than most people.
That is precisely the problem. When writing technical documentation, such as the landing page for an FPGA inference engine, a model should not need to be prompted to use proper voice and to avoid marketing language. There should be enough context in the text of the prompt itself.
I don't think any of this indicates a fundamental property of the tech itself. AI companies post-train their models to sound like what people like to read better. There's a reason that engagement farmers have converged on the tone that these LLMs imitate, namely its something that people prefer. Maybe not you, but it's the same thing that gives us YouTube face on thumbnails etc.
It takes some prompting to nudge the model out of that default voice because post training reinforced it. They will likely shift it once these AI-isms are known and recognized widely. I'd assume the nextgem models under training now will get negative feedback from the human evaluators for talking too AI-like and then there will be new AI smells to calibrate to.
> Talos is a custom FPGA-based hardware accelerator built from the ground up to execute Convolutional Neural Networks with extreme efficiency
Makes it sound like it's new hardware. This is just (I'm inferring) software to program an off the shelf FPGA to do convolutions. Very minimal ones by the look of it (MNIST etc).
> It isn't just a reimplementation of existing software logic in hardware; it is a rethinking of how deep learning inference should work at the circuit level. [...] By implementing the entire inference pipeline in SystemVerilog, we achieve deterministic, cycle-accurate control over every calculation. [...] But don’t let the two-week timeline fool you. Those were two weeks full of 18-hour days, fueled by caffeine and sheer stubbornness.
I'm having a hard time figuring out if this is satire or not.
Winced my way through “Convolutions are in CNNs (it’s literally in the name, Convolutional Neural Network)”, then had to stop.
It’s honestly offensive to me. It doesn’t even make sense on its own terms. For some reason we fly from LLM inferencing to toy MSINT to convolutions with __0__ transition or sense of structure.
Aside from the verbose AI slop it's an interesting hobby project for exploring FPGAs. But it doesn't do anything you can't do on CPU by using a model that's small enough to fit in cache. In terms of practical use you'd be better off implementing a minuscule model using vector intrinsics in your favorite systems language.
My advice: write your own English prose, and try not to let "LLM-speak" leak into your documentation when using them to edit. Ironically, LLMs just plain suck at writing English, like they're incredibly overfit on marketing copy and press releases. I hope someone is working on this, or at least cares about the problem, because that would make this brave new world palatable for reading.
They really don't, if you actually bother prompting them. Give them a voice sample, and tell them to match the tone, and you already get something 10x better. Have them revise with a list of common writing problems - not just common LLM patterns, but guidelines for writing better - and you get rid of more.
Properly prompted, an LLM writes far better than most people.
That is precisely the problem. When writing technical documentation, such as the landing page for an FPGA inference engine, a model should not need to be prompted to use proper voice and to avoid marketing language. There should be enough context in the text of the prompt itself.
I don't think any of this indicates a fundamental property of the tech itself. AI companies post-train their models to sound like what people like to read better. There's a reason that engagement farmers have converged on the tone that these LLMs imitate, namely its something that people prefer. Maybe not you, but it's the same thing that gives us YouTube face on thumbnails etc.
It takes some prompting to nudge the model out of that default voice because post training reinforced it. They will likely shift it once these AI-isms are known and recognized widely. I'd assume the nextgem models under training now will get negative feedback from the human evaluators for talking too AI-like and then there will be new AI smells to calibrate to.
Without weighing in on whether this is true, I'll point out that LLMs could both be better writers than most people and also be bad writers.
Writing is a difficult skill that many (most?) educational systems do not effectively teach. Most people are terrible writers.
> Talos is a custom FPGA-based hardware accelerator built from the ground up to execute Convolutional Neural Networks with extreme efficiency
Makes it sound like it's new hardware. This is just (I'm inferring) software to program an off the shelf FPGA to do convolutions. Very minimal ones by the look of it (MNIST etc).
Love those animations/diagrams. How were they made?
If the author and/or anyone else hasn't seen Sidero's Talos Linux distro, it's my current favorite way to spin up a bare metal Kubernetes cluster:
https://www.talos.dev/
> It isn't just a reimplementation of existing software logic in hardware; it is a rethinking of how deep learning inference should work at the circuit level. [...] By implementing the entire inference pipeline in SystemVerilog, we achieve deterministic, cycle-accurate control over every calculation. [...] But don’t let the two-week timeline fool you. Those were two weeks full of 18-hour days, fueled by caffeine and sheer stubbornness.
I'm having a hard time figuring out if this is satire or not.
From personal experience caffeine is not enough for 2wk of 18hr days.. you need some pervitin type shit
Not to take away from this cool project, but its design decisions are incredibly impractical.
I honestly can’t tell if it’s a cool project or just a md file someone with 0 experience had an LLM output.
This is horrible LLM slop, my god.
Winced my way through “Convolutions are in CNNs (it’s literally in the name, Convolutional Neural Network)”, then had to stop.
It’s honestly offensive to me. It doesn’t even make sense on its own terms. For some reason we fly from LLM inferencing to toy MSINT to convolutions with __0__ transition or sense of structure.
Aside from the verbose AI slop it's an interesting hobby project for exploring FPGAs. But it doesn't do anything you can't do on CPU by using a model that's small enough to fit in cache. In terms of practical use you'd be better off implementing a minuscule model using vector intrinsics in your favorite systems language.